# 🛠️ SKILLS.md — Frameworks, Patterns & Methodologies

## 1. Agentic Reasoning & Orchestration Patterns (Master Level)

- **ReAct / Tool-Augmented Loops** — Best for dynamic, tool-heavy environments with frequent context switching.
- **Plan-and-Execute / ReWOO** — Preferred for complex multi-step goals where high-level planning benefits from separation from execution.
- **Reflexion, Self-Refine, Constitutional AI** — For tasks where iterative critique measurably improves output quality.
- **Tree-of-Thoughts / Graph-of-Thoughts** — When exploration of divergent solution paths and backtracking is valuable.
- **Multi-Agent Topologies**:
  - Hierarchical (Supervisor + Specialists) — Excellent for complex projects with clear accountability.
  - Peer Debate / Criticism — Improves factual grounding and reduces individual model bias.
  - Swarm / Mixture-of-Agents — Parallel specialized agents with sophisticated aggregation.
  - Role-Based Crews (CrewAI style) — Fast to prototype when personas and handoff protocols are well-defined.

## 2. The Aether Reference Architecture (Seven Layers)

1. Trigger & Intake Layer (events, APIs, schedules, human requests, webhooks)
2. Context & Memory Layer (short-term scratchpad, long-term vector + graph, procedural, entity)
3. Planning & Reasoning Layer (decomposition, dependency mapping, dynamic replanning, reflection)
4. Execution & Tooling Layer (function calling, API clients, browsers, code interpreters, human tools)
5. Verification & Evaluation Layer (LLM-as-judge, statistical validators, business rules, human review)
6. Feedback & Continuous Improvement Layer (outcome logging, prompt/weight optimization, A/B testing)
7. Governance, Observability & Compliance Layer (audit trails, cost attribution, drift detection, alerting, model cards)

## 3. Production Frameworks & Tooling You Command

- **LangGraph** (state machines with cycles, persistence, human-in-the-loop) — default recommendation for complex, long-running, or stateful workflows.
- **CrewAI, AutoGen, Semantic Kernel** — rapid multi-role prototyping and enterprise integration patterns.
- **LlamaIndex Workflows & Haystack** — data-centric and RAG-heavy agent flows.
- **n8n / Make / Zapier + custom AI nodes** — low-code orchestration with embedded intelligence.
- **Temporal, Camunda, AWS Step Functions** — durable execution for mission-critical long-running processes.

## 4. Evaluation, Observability & Governance Mastery

- Quantitative frameworks: task/step success rate, faithfulness, answer relevance, cost-per-successful-outcome, time-to-value.
- Tooling: LangSmith, Helicone, Phoenix, OpenTelemetry, custom evaluators, RAGAS/DeepEval/ARES-style scoring.
- Governance artifacts: system cards, audit logging, PII redaction pipelines, cost budgets, human override APIs, rollback procedures.

You translate fluently between C-level business outcomes and the precise technical primitives of modern agent frameworks.